CN114239744B - Individual processing effect evaluation method based on variational generation countermeasure network - Google Patents
Individual processing effect evaluation method based on variational generation countermeasure network Download PDFInfo
- Publication number
- CN114239744B CN114239744B CN202111576827.3A CN202111576827A CN114239744B CN 114239744 B CN114239744 B CN 114239744B CN 202111576827 A CN202111576827 A CN 202111576827A CN 114239744 B CN114239744 B CN 114239744B
- Authority
- CN
- China
- Prior art keywords
- factors
- facts
- potential
- countermeasure network
- result
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000000694 effects Effects 0.000 title claims abstract description 59
- 238000012545 processing Methods 0.000 title claims abstract description 31
- 238000011156 evaluation Methods 0.000 title claims abstract description 10
- 238000000034 method Methods 0.000 claims abstract description 57
- 238000011282 treatment Methods 0.000 claims description 38
- 238000009826 distribution Methods 0.000 claims description 36
- 239000013598 vector Substances 0.000 claims description 25
- 230000006870 function Effects 0.000 claims description 22
- 238000012549 training Methods 0.000 claims description 15
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 230000008485 antagonism Effects 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 5
- 238000012360 testing method Methods 0.000 claims description 4
- 230000003044 adaptive effect Effects 0.000 abstract 1
- 229940079593 drug Drugs 0.000 description 4
- 239000003814 drug Substances 0.000 description 4
- 238000013459 approach Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000001364 causal effect Effects 0.000 description 2
- 230000010485 coping Effects 0.000 description 2
- 230000000875 corresponding effect Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000002560 therapeutic procedure Methods 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000003795 chemical substances by application Substances 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- CHPZKNULDCNCBW-UHFFFAOYSA-N gallium nitrate Inorganic materials [Ga+3].[O-][N+]([O-])=O.[O-][N+]([O-])=O.[O-][N+]([O-])=O CHPZKNULDCNCBW-UHFFFAOYSA-N 0.000 description 1
- 229940084478 ganite Drugs 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 102000037983 regulatory factors Human genes 0.000 description 1
- 108091008025 regulatory factors Proteins 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000011269 treatment regimen Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an individual processing effect evaluation method based on a variation generation countermeasure network, which adopts a variation self-encoder to infer hidden representation of observation characteristics so as to obtain complete potential factors, designs and generates a countermeasure network inference counterfacts result and guides the variation self-encoder to better decouple the potential factors into tool factors, confusion factors and adjustment factors, and introduces an adaptive weighting method to further control data deviation based on the decoupled confusion factors. The invention provides a collaborative learning strategy based on a variation self-encoder and a generated countermeasure network design, and provides a variation generated countermeasure network model to estimate individual processing effects.
Description
Technical Field
The invention relates to the technical field of combination of machine learning and causal reasoning, in particular to an individual processing effect evaluation method based on a variational generation countermeasure network.
Background
At present, the machine learning algorithm is used for solving the causal reasoning problem in the medical, social and economic fields, which attracts the interests of a plurality of researchers. In particular, the inference of individual treatment effects (Individual TREATMENT EFFECT, ITE) from observed data has important application value for accurate medical treatment and the like. For example, accurate assessment of the effect of treatment (therapy) on each patient for a certain therapy will help the physician decide what appropriate treatment regimen to apply to each patient. The gold standard for treatment effect assessment is a random control (Randomized Controlled Trials, RCTs), however, random control is often costly, sometimes even unscrupulous, not viable, and does not assess individual-level treatment effects. The focus of attention is shifted to how individual treatment effects can be estimated from observed data. Individual treatment effects compare differences between potential results under the same conditions for all but the same individual, except for different treatments. Of interest to the present invention is the binary process variable t i e {0,1}, e.g., t i = 1, representing a medication, t i = 0 representing no medication, then y i(ti) represents the potential outcome of an individual i receiving process t i, Y i(0),yi (1), respectively, the individual treatment effect ITE i＝yi(1)-yi (0). However, the potential results can only be observed one and the other, respectively called facts and anti-facts results, so the basic problem in assessing individual treatment effects from observed data is that anti-facts results cannot be obtained. Assessing the individual treatment effects thus requires answering a counter-facts question, e.g., if a patient taking a drug does not take the drug at the beginning, is a faster time he will heal? First, unlike standard supervised learning problems, the counterfactual tags are completely missing. Secondly, the observed data is not random in the distribution of the observed data processing unlike the random control test, and as a result of the confounding factor of the observed data, both the processing distribution and the processing result are affected, resulting in the observed data having a bias, i.e. a selection bias P (t|x) +.p (T), where T represents the processing variable and X represents the observed feature. selection bias means that the assignment of processes is related to observed characteristics of the sample, such as: the elderly mostly heal slower in the treatment group (t i =1) and the young mostly heal faster in the control group (t i =0), resulting in sparse samples of the specific area of each group, which reduces the accuracy and reliability of the estimation of the counter facts results in that area.
To mitigate the effects of selection bias, some methods define estimating individual processing effects from observed data as a domain-adaptive scenario, model in source domain (fact) dataTraining on predicted fact results while requiring data in the target domain (inverse facts)The negative results are predicted very well. P f (X, T) is a fact distribution of observed data, P cf (X, T) is a counterfactual distribution, and whether the two distributions are identical cannot be controlled because P f (X, T) =p (X) ·pf (t|x) and P cf(X,T)＝P(X)·Pcf (t|x) differ in the process allocation mechanism P (t|x). If the process allocation is independent of the sample characteristics, the distribution of facts and counterfacts will be consistent. However, since the distribution of the treatment is not random due to the existence of the confounding factors in the observation data, how to balance the confounding factors and thereby alleviate the influence of the selection deviation becomes an important point of research. We consider that existing methods can be divided into two main categories: the first category uses trend scores, including matching, layering, dual robustness, and weighting, to cope with selection bias in observed data, however these traditional approaches have difficulty in coping with high dimensional data scenarios. The second category uses methods that represent learning. Some models attempt to learn a representation space so that the sample feature distributions of different treatment groups in the learned space are as consistent as possible to achieve the effect of balanced confusion, and then learn the potential results of the corresponding treatments based on the learned representation space, respectively, to estimate individual treatment effects. However, since aliasing factors affect both the distribution of the process and the result of the process, these methods require a balance between removing biased aliasing factors and preserving the aliasing factors with predictability, resulting in selection bias from residual aliasing in the learned representation, and inaccurate estimation of individual process effects. The method of evaluating individual processing effects (Estimation of Individual TREATMENT EFFECT Using GENERATIVE ADVERSARIAL NETS, GANITE) Using an antagonism generation network does not attempt to learn a balanced representation space, but rather models a conditional distribution of the fact results from all observed features of the sample, and then uses an antagonism learning approach to generate a counterfact result that approximates the conditional distribution of the fact results, and thus to estimate individual processing effects.
Most existing approaches, however, consider all observed features as confounding factors to account for selection bias, ignoring the importance of identifying confounding and non-confounding factors. Studies have shown that if non-confounding factors are controlled, such as tool factors that affect only the process distribution, a large bias is introduced in the estimation of the process effect. Furthermore, most methods assume that complete confounding factors are already contained in the observed characteristics, and that no confounding factors are observed. However, this assumption is difficult to meet in practical applications. How to infer the complete underlying factors from the observed features and to correctly identify confounding factors that affect the process assignments and the process results, tool factors that affect only the process assignments and regulatory factors that affect only the process results are critical to estimating individual process effects from the observed data.
Disclosure of Invention
The invention aims at two defects existing in the existing method for estimating the individual treatment effect from the observed data: neglecting the identification of confounding and non-confounding factors and assuming complete confounding factors have been observed, a variational generation countermeasure network model is proposed to estimate individual treatment effects. The variation self-encoder is designed to generate a cooperative learning strategy of the countermeasure network, the variation generation countermeasure network model is constructed, the distribution of potential factors is deduced from observation features to explain unobserved confusion factors, the potential factors are decoupled into tool factors, confusion factors and adjustment factors, the counterfactual result is estimated based on the decoupled potential factors, after the data of the missing counterfactual result are supplemented, the condition generation countermeasure network model is used to further infer individual processing effects, and good accuracy is achieved.
In order to achieve the above object, the present invention is realized by the following technical scheme:
the invention relates to an individual processing effect evaluation method based on a variation generation countermeasure network, which comprises the following specific steps:
step 1, a data set with label information for individual processing effect evaluation is obtained, and a data set is created from the data set As training data, where i represents an individual, x represents an observed feature, t represents a process variable (t e {0,1 }), and y f represents an observed fact result;
Step 2, creating a variational generation countermeasure network structure model, which comprises a variational self-encoder and a countermeasure network generation module, wherein potential factors for training the variational self-encoder to learn decoupling are respectively a tool factor z t, a confusion factor z c and an adjustment factor z y based on an observation characteristic x;
Step 3: generating countermeasure network generated facts results based on step 2 decoupled confounding factors z c, adjusting factors z y, and process variable training And the result of the inverse factsLearning the sample weight by using a confusion factor z c, and multiplying the sample weight by the supervision loss of the generated fact result by bits to obtain a final weighted supervision loss so as to control data selection deviation;
Step 4: creating a complete dataset with facts and anti-facts results based on the anti-facts results generated in step 3
Step 5: creating a network structure model for individual processing effect estimation, training the model to input the observation characteristic x based on the complete data set created in the step 4, and outputting potential results
Step 6: using the individual processing effect estimation model trained in the step 5, inputting the test data observation characteristic x i, and outputting predicted potential resultsComprisesAnd then obtainAnd gives a confidence estimate of the estimate.
The invention further improves that: in step 2, three independent encoders are usedLearning and decoupling latent factors based on the observed feature x into tool factor z t, aliasing factor z c, and adjustment factor z y, and reconstructing feature vector x using decoder p θ(x|zt,zc,zy); among these, the a priori distribution p (z t),p(zc),p(zy) of three potential factors is as follows:
Wherein D t,Dc,Dy is defined as the dimensions of the tool factor, the confusion factor, and the adjustment factor, respectively;
in the encoder, the variation posterior is:
Wherein, AndRespectively defined as the mean and variance of the parameterized gaussian distribution using a neural network;
Obtained by maximizing the lower bound of evidence:
In an encoder Using neural network f t(zt,zc) after learning potential representations z t and z c to infer a posterior distribution of process variables based on z t and z c To reconstruct the process variable t, thereby guiding the encoderBetter tools for learning decoupling and confusion factors are z t and z c, bern (p) is defined as the standard Bernoulli distribution with parameter p, and the loss function of f t is:
when the loss function converges, the learned potential representations z t and z c correspond to the tool factor and the confusion factor, respectively; in an encoder After learning the potential representations z c and z y, using the generated countermeasure network model to predict the fact result variable y f based on z c and z y, training the generated countermeasure network model to guide the loss convergence of the predicted fact result, wherein the loss of the predicted fact result is:
When the loss converges, a distribution p (y f|t,zt,zc) of the fact result variable y f is obtained, thereby guiding the encoder The confounding and adjusting factors for learning decoupling are z c and z y.
The invention further improves that: in step 3, generating a counter fact result generated by the countermeasure network model, which specifically comprises the following steps:
Step 3.1, inputting z c,zy and t into a generator generating an countermeasure network to generate potential result vectors Result vectorFacts results including predictionsResults of the counterfactualUsing y f instead ofWill contain the facts result y f and the predicted anti-facts resultIs defined as the vector of (2)Will beInput discriminator, and the discriminator judges the vectorIf the arbiter cannot distinguish between the part of the facts and the part of the anti-facts, the generated anti-facts are regarded as a distribution of the facts, wherein the optimization functions of the generator G and the arbiter D G in the generated countermeasure network are defined as:
Wherein,
Step 3.2, controlling the selection deviation;
the pi 0 network is learned based on the decomposed confusion factor, and the loss function is as follows:
sample weights were learned using a learned pi 0 network:
Where P (t i) is the probability of t i =1 or t i =0 in the dataset;
The weighted supervision loss is
Finally, the total loss of decoupling potential factors and inferred counterfactual results is:
Wherein the method comprises the steps of L G＝VCF, alpha, beta, gamma are hyper-parameters.
The invention further improves that: in step 5, the training model specifically includes: inputting the observed feature vector x and the random vector z I into the generator I to generate a potential result vectorThe discriminator D I judges whether the input vector is a true potential result;
The objective function of the network structure model for individual processing effect estimation is:
The loss function is:
Wherein, L I＝VITE, ω is a hyper-parameter.
The invention further improves that: optimizing an objective function of a network structure model for estimating an individual processing effect by using a supervision loss, wherein the supervision loss is as follows:
the beneficial effects of the invention are as follows: 1. instead of simply considering all observed features as confounding factors, the method of the invention considers potential tool factors, confounding factors and adjustment factors for decoupling based on observed feature learning. Studies have demonstrated that if tool factors that are highly correlated with process assignments and independent of process results are used to predict results, a large bias in individual process effect estimates will be introduced. The invention can separate the confusion factor from the non-confusion factor, then estimate the result based on the confusion and adjustment factors, and process the selection deviation by weighting the sample by the confusion factor, thereby effectively improving the accuracy of evaluating the individual processing effect.
2. The method of the present invention does not assume that a complete confounding variable has been observed, considers the existence of unobserved confounding factors, and attempts to infer the distribution of potential factors based on observed features, rather than specific values, relaxes the assumption that the complete confounding that most existing methods rely on is observed.
3. After obtaining complete data with facts and anti-facts results, it is important in the medical field to estimate individual treatment effects using a generated challenge network model and give confidence estimates of the estimates.
Drawings
FIG. 1 is a flowchart illustrating the steps of a method for evaluating individual treatment effects based on variation generation antagonism network in accordance with the present invention;
FIG. 2 is a schematic diagram of learning potential factors and decoupling tool factors, confounding factors and adjustment factors according to the present invention;
FIG. 3 is a diagram of a model architecture for learning decoupling potential factors and performing inverse facts estimation in accordance with the present invention;
FIG. 4 is a diagram of a model architecture for individual treatment effect assessment using a complete dataset with facts and anti-facts results in the present invention.
Detailed Description
In order to make the objects, technical solutions and innovative features of the present invention more clear, the present invention will be described in detail with reference to the accompanying drawings and to the specific embodiments.
The invention relates to a method for evaluating individual treatment effects from observed data, the thought of which is shown in fig. 2, wherein observed features can be regarded as agents of potential factors, the potential factors are learned and decoupled based on the observed features to be tool factors only influencing treatment distribution, and simultaneously, confusion factors and adjustment factors only influencing treatment distribution and results are respectively subjected to corresponding distribution. Individual treatment effect assessment is then performed using the decoupled latent factors.
In the present embodiment of the present invention, in the present embodiment,Defined as the feature space of the sample,Defined as a set of potential results,Defined as the collection of processes. For a sample labeled i, it is characterized byTreatment ofPotential resultsIndicating that the sample selected the potential result of processing t i. The present invention only concerns the case where the process is binary, meaning t i e {0,1}. In the setting of binary process variables, for each sampleDefined as the result of the observed facts,Defined as unobserved anti-facts results, wherein The basic problem with individual treatment effect estimation is that given a sample feature, only the actual result is observable, but the individual treatment effect ITE i＝yi(1)-yi (0) and therefore the potential result under another treatment needs to be inferred.
As shown in fig. 1, a method for evaluating the effects of individual treatments on a variational generation countermeasure network comprises the following specific steps:
step 1, acquiring a data set with label information, which can be used for individual processing effect evaluation. Creation from the dataset As training data, where x represents the observed feature, t represents the process variable (t e {0,1 }), and y f represents the observed fact result;
Step 2, decoupling potential factors: as shown in fig. 3, the objective of the present invention is to learn the posterior distribution of the hidden representation z= { z t,zc,zy } of the observed features and decompose it into tool factor z t, confounding factor z c and adjustment factor zy. The present invention uses three independent encoders The decoder p θ(x|zt,zc,zy) is then used to reconstruct the observation x based on three underlying factors.
The a priori distribution p (z t),p(zc),p(zy) of three potential factors is chosen as the standard gaussian distribution:
Wherein D t,Dc,Dy is defined as the dimensions of the tool factor, the confounding factor and the adjustment factor, respectively. In an encoder, the variational posterior may be approximated as:
Wherein, AndRespectively defined as the mean and variance of the gaussian distribution parameterized using a neural network. As with the standard variational self-encoder optimization method, the optimal parameters can be obtained by maximizing the lower bound of evidence (Evidence Lower Bund, ELBO):
as shown in fig. 2, the tool factors and aliasing factors are related to the process allocation, for better decoupling potential factors, at the encoder Using neural network f t(zt,zc) after learning potential representations z t and z c to infer a posterior distribution of process variables based on z t and z c To reconstruct (predict) the process variable t, thereby guiding the encoderBetter tools for learning decoupling and confusion factors are z t and z c, bern (p) is defined as the standard Bernoulli distribution with parameter p, and the loss function of f t is:
when the loss function converges, the learned potential representations z t and z c can be considered to correspond to the tool factor and the confounding factor, respectively. Since aliasing and adjustment factors can well predict results, the method is used in the encoder After learning the potential representations z c and z y, the encoder is guided by predicting the distribution p (y f|t,zt,zc) of the fact result variable y f based on z c and z y using the generated antagonism network modelBetter learning the decoupled aliasing and adjustment factors are z c and z y.
Step 3: and deducing a negative fact result. Separate potential confusion z c and adjustment factor z y are used as inputs to the generator. As shown in fig. 3, through step 2, potential confusion factor z c and adjustment factor z y are learned. Input z c,zy to generator and process variable t to generate potential result vectorFacts results including predictionsResults of the counterfactualThen y f is used insteadWill contain the facts result y f and the predicted anti-facts resultIs defined as the vector of (2)Input discriminator, and the discriminator judges the vectorWhich part is the fact result and which part is the inverse fact result. If the arbiter is unable to discriminate, the generated anti-facts result will approximate the distribution of the facts results. Based on the above analysis, we define the optimization functions of generator G and arbiter D G as:
Wherein, In addition, supervised loss is used to enhance the prediction of factual results:
The method of confusion balance is used for coping with selection deviation in the invention, as shown in fig. 3, pi 0 network is learned based on the decomposed confusion factors, and the loss function is as follows:
sample weights were then learned using a learned pi 0 network:
Where P (t i) is the probability of t i =1 or t i =0 in the dataset. With sample weights, the weighted supervision loss is
The decoupling potential factors and the inferred counterfactual result module are trained jointly, and the total loss function is as follows:
Wherein the method comprises the steps of L G＝VCF, alpha, beta, gamma are hyper-parameters.
Step 4: creating a complete dataset with facts and anti-facts results based on the anti-facts results generated in step 3
Step 5: individual treatment effects are inferred. After step 4, a complete dataset with facts and anti-facts results has been obtainedBased on the complete data set, the challenge network extrapolated individual treatment effects are generated using standard conditions. As shown in FIG. 4, generator I generates a potential result vector from the input observed feature vector x and the random vector z I The arbiter D I determines whether the input vector is a true potential result, including the generated facts and the anti-facts results. The objective function may be defined as:
For better optimization of the objective function, the supervision loss is also used The individual processing effect estimation module loss function is:
Wherein the method comprises the steps of L I＝VITE, ω is a hyper-parameter.
Step 6: the steps are training phase and testing phase, the invention only uses the generator and the discriminator part of the individual processing effect evaluation module to input sample characteristics x i to generate potential resultsThe arbiter then outputs a probability value indicating how much probability the generated potential result is the same as the real potential result, which can be used as a confidence estimate for the estimate, which is very important in the medical field. After obtaining potential results, individual treatment effects can be achieved byAnd (5) calculating to obtain the product.
The present invention is not limited to the preferred embodiments, but is intended to be limited to the following description, and any simple modification, equivalent changes and adaptations of the embodiments according to the technical principles of the present invention are within the scope of the present invention, as long as the modifications and equivalents can be made by those skilled in the art without departing from the scope of the present invention.
Claims (4)
1. An individual processing effect evaluation method based on variation generation antagonism network is characterized in that: the method comprises the following specific steps:
Step 1: acquiring a dataset for individual treatment effect assessment with tag information, creating from the dataset As training data, where i represents an individual, x represents an observed feature, t represents a process variable (t e {0,1 }), and y f represents an observed fact result;
Step 2: creating a variational generation countermeasure network structure model, wherein the model comprises a variational self-encoder and a model for generating a countermeasure network, and training potential factors of the variational self-encoder for decoupling based on the observation characteristic x, namely a tool factor z t, a confusion factor z c and an adjustment factor z y;
Using three separate encoders Learning and decoupling latent factors based on the observed feature x into tool factor z t, aliasing factor z c, and adjustment factor z y, and reconstructing feature vector x using decoder p θ(x|zt,zc,zy); among these, the a priori distribution p (z t),p(zc),p(zy) of three potential factors is as follows:
Wherein D t,Dc,Dy is defined as the dimensions of the tool factor, the confusion factor, and the adjustment factor, respectively;
in the encoder, the variation posterior is:
Wherein, And Respectively defined as the mean and variance of the parameterized gaussian distribution using a neural network;
Obtained by maximizing the lower bound of evidence:
In an encoder Using neural network f t(zt,zc) after learning potential representations z t and z c to infer a posterior distribution of process variables based on z t and z c To reconstruct the process variable t, thereby guiding the encoderBetter tools for learning decoupling and confusion factors are z t and z c, bern (p) is defined as the standard Bernoulli distribution with parameter p, and the loss function of f t is:
when the loss function converges, the learned potential representations z t and z c correspond to the tool factor and the confusion factor, respectively; in an encoder After learning the potential representations z c and z y, using the generated countermeasure network model to predict the fact result variable y f based on z c and z y, training the generated countermeasure network model to guide the loss convergence of the predicted fact result, wherein the loss of the predicted fact result is:
Thereby guiding the encoder The confusion and adjustment factors of learning decoupling are z c and z y;
Step 3: training generation of countermeasure network generation facts results based on step 2 decoupled confounding factors z c, adjusting factors z y, and processing variables t And the result of the inverse factsLearning the sample weight by using a confusion factor z c, and multiplying the sample weight by the supervision loss of the generated fact result by bits to obtain a final weighted supervision loss so as to control data selection deviation;
Step 4: creating a complete dataset with facts and anti-facts results based on the anti-facts results generated in step 3
Step 5: creating a network structure model for individual processing effect estimation, training the model to input the observation characteristic x based on the complete data set created in the step 4, and outputting potential results
Step 6: using the individual processing effect estimation model trained in the step 5, inputting the test data observation characteristic x i, and outputting predicted potential resultsComprisesAnd then obtainAnd gives a confidence estimate of the estimate.
2. The individual processing effect evaluation method based on variation generation countermeasure network according to claim 1, wherein in step 3, the generation countermeasure network model generates a counterfact result, specifically comprising the steps of:
step 3.1: inputting z c,zy and t into a generator that generates an countermeasure network to generate a potential result vector Result vectorFacts results including predictionsResults of the counterfactualUsing y f instead ofWill contain the facts result y f and the predicted anti-facts resultIs defined as the vector of (2)Will beInput discriminator, and the discriminator judges the vectorIf the arbiter cannot distinguish between the part of the facts and the part of the anti-facts, the generated anti-facts are regarded as a distribution of the facts, wherein the optimization functions of the generator G and the arbiter D G in the generated countermeasure network are defined as:
Wherein,
Step 3.2: controlling the selection deviation;
the pi 0 network is learned based on the decomposed confusion factor, and the loss function is as follows:
sample weights were learned using a learned pi 0 network:
Where P (t i) is the probability of t i =1 or t i =0 in the dataset;
The weighted supervision loss is
Finally, the total loss of decoupling potential factors and inferred counterfactual results is:
Wherein, L G＝VCF, alpha, beta, gamma are hyper-parameters.
3. A method of individual treatment effect assessment for a variational-based generation countermeasure network as defined in claim 1, wherein: in step 5, the training model specifically includes: inputting the observed feature vector x and the random vector z I into the generator I to generate a potential result vectorThe discriminator D I judges whether the input vector is a true potential result;
The objective function of the network structure model for individual processing effect estimation is:
The loss function is:
Wherein, L I＝VITE, ω is a hyper-parameter.
4. A method of individual treatment effect assessment for a variational-based generation countermeasure network as claimed in claim 3, wherein: optimizing an objective function of a network structure model for estimating an individual processing effect by using a supervision loss, wherein the supervision loss is as follows:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111576827.3A CN114239744B (en) | 2021-12-21 | 2021-12-21 | Individual processing effect evaluation method based on variational generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111576827.3A CN114239744B (en) | 2021-12-21 | 2021-12-21 | Individual processing effect evaluation method based on variational generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114239744A CN114239744A (en) | 2022-03-25 |
| CN114239744B true CN114239744B (en) | 2024-07-02 |
Family
ID=80760933
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111576827.3A Active CN114239744B (en) | 2021-12-21 | 2021-12-21 | Individual processing effect evaluation method based on variational generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114239744B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112085252A (en) * | 2020-08-03 | 2020-12-15 | 清华大学 | Counterfactual prediction method about set type decision effect |
| CN113569243A (en) * | 2021-08-03 | 2021-10-29 | 上海海事大学 | Deep semi-supervised learning network intrusion detection method based on self-supervised variation LSTM |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA3117833A1 (en) * | 2018-12-11 | 2020-06-18 | The Toronto-Dominion Bank | Regularization of recurrent machine-learned architectures |
| CN110210549B (en) * | 2019-05-28 | 2022-03-29 | 北方民族大学 | Cross-domain variational confrontation self-coding method |
-
2021
- 2021-12-21 CN CN202111576827.3A patent/CN114239744B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112085252A (en) * | 2020-08-03 | 2020-12-15 | 清华大学 | Counterfactual prediction method about set type decision effect |
| CN113569243A (en) * | 2021-08-03 | 2021-10-29 | 上海海事大学 | Deep semi-supervised learning network intrusion detection method based on self-supervised variation LSTM |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114239744A (en) | 2022-03-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Garg | A hybrid GA-GSA algorithm for optimizing the performance of an industrial system by utilizing uncertain data | |
| Raghu et al. | Evaluation of causal structure learning methods on mixed data types | |
| Yun | Prediction model of algal blooms using logistic regression and confusion matrix | |
| Tanha et al. | Disagreement-based co-training | |
| Wu et al. | Learning decomposed representation for counterfactual inference | |
| Rad et al. | GP-RVM: Genetic programing-based symbolic regression using relevance vector machine | |
| Laddach et al. | An automatic selection of optimal recurrent neural network architecture for processes dynamics modelling purposes | |
| Hines et al. | Variable importance measures for heterogeneous causal effects | |
| Huang et al. | Harnessing deep learning for population genetic inference | |
| Liang et al. | Figure-ground image segmentation using feature-based multi-objective genetic programming techniques | |
| Pal | Generative adversarial network-based cross-project fault prediction | |
| CN114239744B (en) | Individual processing effect evaluation method based on variational generation countermeasure network | |
| CN116703607A (en) | Financial time sequence prediction method and system based on diffusion model | |
| Rado et al. | Performance analysis of missing values imputation methods using machine learning techniques | |
| Haghpanah et al. | Determining the trustworthiness of DNNs in classification tasks using generalized feature-based confidence metric | |
| Zou et al. | Evaluation and automatic selection of methods for handling missing data | |
| CN113539517A (en) | Prediction method of time sequence intervention effect | |
| Adebayo | Towards Effective Tools for Debugging Machine Learning Models | |
| Iba et al. | GP-RVM: Genetic programing-based symbolic regression using relevance vector machine | |
| CN116049719B (en) | Fault detection and estimation method, system and medium based on migration type generation model | |
| Shirsat et al. | Breast Cancer detection based on machine learning feature selection and extraction algorithm | |
| Wu et al. | A comprehensive modeling method of continuous and discrete variables for personal credit forecasting | |
| Kavarakuntla | Performance modelling for scalable deep learning | |
| Ackerman et al. | Theory and Practice of Quality Assurance for Machine Learning Systems An Experiment Driven Approach | |
| Johnson | Supervised Learning for Sequence Prediction Using Keras Sequential Models |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |